Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!uunet!mcsun!ukc!edcastle!aiai!richard
From: richard@aiai.ed.ac.uk (Richard Tobin)
Newsgroups: comp.arch
Subject: Re: ~8-job "knee" in response curves on Suns (was Re: IBM RS6000)
Message-ID: <3956@skye.ed.ac.uk>
Date: 16 Jan 91 19:28:27 GMT
References: <1991Jan10.214122.9506@news.arc.nasa.gov> <amos.663857722@shum> <5257@auspex.auspex.com>
Reply-To: richard@aiai.UUCP (Richard Tobin)
Organization: AIAI, University of Edinburgh, Scotland
Lines: 199

I've several times heard about the "knee" in the performance of Suns as
the number of concurrent processes increases, so I wrote a simple
benchmark.

It forks n processes, which then write to each other in a circle of
pipes.  Note that all but one process is blocked waiting for input
at any time.

Here are some results.  The program is at the end.

Sun 3/75:

The knee is noticeable.  Switching rate is about constant for up to 7
processes, then it starts to decline.

2 : 897 switches / second
3 : 862 switches / second
4 : 860 switches / second
5 : 856 switches / second
6 : 851 switches / second
7 : 860 switches / second
8 : 686 switches / second
9 : 650 switches / second
10 : 641 switches / second
11 : 621 switches / second
12 : 621 switches / second
13 : 589 switches / second
14 : 597 switches / second


Sun 3/260:

Similar but more severe.

2 : 1626 switches / second
3 : 1648 switches / second
4 : 1587 switches / second
5 : 1543 switches / second
6 : 1474 switches / second
7 : 1431 switches / second
8 : 876 switches / second
9 : 608 switches / second
10 : 632 switches / second
11 : 618 switches / second
12 : 598 switches / second
13 : 560 switches / second
14 : 577 switches / second


Sparcstation 1:

Rather variable, gets very slow above 8 processes.

2 : 1971 switches / second
3 : 1944 switches / second
4 : 1944 switches / second
5 : 1092 switches / second
6 : 1892 switches / second
7 : 1184 switches / second
8 : 835 switches / second
9 : 416 switches / second
10 : 396 switches / second
11 : 396 switches / second
12 : 379 switches / second
13 : 359 switches / second


Sun 4/470:

No obvious knee.

2 : 4228 switches / second
4 : 4537 switches / second
6 : 4370 switches / second
8 : 4215 switches / second
10 : 4134 switches / second
12 : 3807 switches / second
14 : 4020 switches / second
16 : 4013 switches / second
18 : 3998 switches / second
20 : 3922 switches / second
22 : 3465 switches / second
24 : 3737 switches / second
26 : 3541 switches / second
28 : 3448 switches / second
30 : 3644 switches / second
32 : 3598 switches / second
34 : 3563 switches / second
36 : 3500 switches / second


-- Richard

/* see how fast we can switch between processes */

#include <stdio.h>
#include <sys/time.h>

int master = 1;
int from, to;

main(argc, argv)
int argc;
char **argv;
{
    int one_to_two[2], n_to_1[2];
    int iter, procs, reads = 0, pid;
    static char buf[] = "x";
    struct timeval tp1, tp2;

    procs = atoi(argv[1]);
    iter = argc > 2 ? atoi(argv[2]) : 100;

    pipe(one_to_two);
    pipe(n_to_1);

    from = n_to_1[0];
    to = one_to_two[1];

    make_child(procs-1, one_to_two, n_to_1[1], n_to_1[0]);

    pid = getpid();

    if(master)
	write(to, buf, 1);

    gettimeofday(&tp1, (struct timezone *)0);

    while(1)
    {
	switch(read(from, buf, 1))
	{
	  case -1:
	    perror("read");
	    exit(1);

	  case 0:
#ifdef DEBUG
	    printf("process %d exiting after %d reads\n", pid, reads);
#endif
	    if(master)
	    {
		double sec;

		gettimeofday(&tp2, (struct timezone *)0);
		sec = (tp2.tv_sec - tp1.tv_sec) +
		      (tp2.tv_usec - tp1.tv_usec) / 1e6;
		printf("%d switches / second\n", (int)(procs*iter/sec + 0.5));
	    }
	    exit(0);

	  case 1:
	    reads++;
#ifdef DEBUG
	    printf("process %d read %c\n", pid, buf[0]);
#endif
	    if(master && reads == iter)
		close(to);
	    else
		write(to, buf, 1);
	}
    }
}

make_child(n, parentpipe, lastwriter, closeme)
int n, parentpipe[2], lastwriter, closeme;
{
    if(fork() == 0)
    {
	/* child */

	master = 0;

	close(parentpipe[1]);
	close(closeme);
	from = parentpipe[0];

	if(n == 1)
	    /* This is the last process */
	    to = lastwriter;
	else
	{
	    int p[2];

	    pipe(p);
	    to = p[1];
	    make_child(n-1, p, lastwriter, from);
	}
    }
    else
    {
	close(parentpipe[0]);
	close(lastwriter);
    }
}
-- 
Richard Tobin,                       JANET: R.Tobin@uk.ac.ed             
AI Applications Institute,           ARPA:  R.Tobin%uk.ac.ed@nsfnet-relay.ac.uk
Edinburgh University.                UUCP:  ...!ukc!ed.ac.uk!R.Tobin